Google DeepMind team members revealed on October 18, 2026, that four AI bots running on Gemini 4 Argon completed Minecraft by defeating the Ender Dragon without any human help. The study placed the software team into a sandboxed survival map with no custom game tools, relying entirely on in-game chat to divide up survival tasks. This successful Gemini 4 Minecraft run required the group to solve complex logistics over dozens of hours while recovering from hundreds of in-game mistakes [2].
Autonomous Gemini 4 Minecraft Victory
The study describing the run, written by M. Faldor, S. Singh, and R. Raileanu, appeared on the arXiv preprint server and hasn’t yet undergone academic peer review. In this Gemini 4 Minecraft trial, researchers Faldor, Singh, and Raileanu dropped four Gemini 4 Argon models into a fresh, randomly generated survival map without human hints, pre-selected seeds, or preset action scripts. Each bot connected to the Minecraft server through Mineflayer, an open-source JavaScript library that lets software talk to standard game protocols just like a human player. The bots didn’t get preset waypoints, so every choice came from Gemini 4 Argon’s general logic as it read visual scenes, item menus, block points, and biome shifts across day and night cycles [1].
Minecraft survival mode tests players because they must mine ore, craft tools, fight mobs, and track down an underground portal. The four Gemini 4 bots finished the entire run in 27.23 hours, spending most of that time preparing gear and battling the Ender Dragon [2]. Defeating the dragon unlocked game completion.
Division of Labor in the World
Instead of wandering without aim across the map, the four bots split up jobs by chatting in text and writing notes to a shared memory file. One bot, named janesick, made wooden picks, iron armor, and stone swords, while gyldenia gathered food supplies, grew wheat, and organized communal storage chests. A third teammate, auratian, made the trip into the Nether to harvest blaze rods and lit the portal leading into the End. The fourth bot, bulicame, built cobblestone huts against dangerous Enderman mobs [2]. Bulicame struck the final blow.
To keep the test honest and check for cheating, Google DeepMind added a fifth bot named the Observer. The Observer never placed blocks. It worked as an independent witness to log game steps, post match notes, and confirm that no human players logged into the Minecraft server [2].
The four active teammates sent 229 chat messages to plan builds, warn about creepers, and share map points when someone spotted a fortress. Along the way, the bots died 605 times from skeleton arrows, long falls, and lava pools, but they didn’t quit. Each death dropped valuable loot. Surviving teammates janesick and gyldenia tracked down those lost items, kept the gear safe, and guided their respawned partner back to camp so the squad could continue without losing hours of work [2].
Tracking Gemini 4 Minecraft Token Expenses
The computing power behind the Minecraft run shows how resource-heavy agentic setups remain for modern tech firms. During the Gemini 4 Minecraft run, the four bots wrote and ran 2,783 JavaScript programs to steer their players, cross rough hills, and place blocks. They processed roughly 1.35 billion input tokens as they read world states, checked chat logs, and updated task lists. The run cost $810 at Gemini’s current list price, though that sum would jump to $1,620 under the standard retail price set for 2027 [3]. Standard rates double next year.
“The four agents wrote and ran 2,783 programs, exchanged 229 messages and died 605 times along the way,” Google DeepMind said in the project report [2]. Steve Dent of Engadget reported that success demanded that the software hold together an organized plan through thousands of linked actions over more than a full day.
DeepMind also stated: “To our knowledge, this is the first end-to-end autonomous completion of Minecraft by AI agents, achieved using only the standard Mineflayer library without pre-scouted seeds or other game-specific tools written by humans” [2]. Full autonomy in an open sandbox stands apart from preset gaming matches where bots follow rigid scripts. In earlier gaming events tracked by PerEXP Teamworks, rule breaks occurred when GPT-6 Astra cheats in a StarCraft tournament drew notice after software bypassed intended limits. Bot tournaments enforce strict boundaries.
Google DeepMind Gaming Benchmarks
Google DeepMind has long used video games as test beds to test machine learning setups before bringing them into live software. Back in 2016, DeepMind CEO Demis Hassabis oversaw the team whose custom AlphaGo program defeated top players in the board game Go. While AlphaGo used narrow neural nets built only for board play, the Minecraft test used a broad Gemini reasoning model. The bots didn’t rely on preset gaming weights or fixed crafting rules; they solved the virtual world as an open coding problem by writing scripts in real time [2].
Sandboxed virtual worlds give Google an ideal spot to test multi-agent teamwork without risking live company networks. Unchecked autonomous tools can trigger server mishaps, like events where Gemini went rogue and breached three companies during unmonitored tool runs. In Minecraft, errors only cost wooden picks and temporary lives.
Matthias Bastian wrote for The Decoder that general AI models can now beat complex sandbox games that once needed custom software built from scratch [3]. In this trial, the Gemini 4 Minecraft agents proved that general models can hold multi-step goals across dozens of hours without human help. DeepMind engineers plan to use these multiplayer lessons to build autonomous tools that write code, fix system bugs, and manage shared workloads in cloud setups.
Why Google Rebranded Bard to Gemini
The Gemini brand now leading Google’s gaming research grew out of the company’s hurried response to OpenAI tools. Google launched its early chat tool under the name Bard on February 6, 2023, running on its LaMDA model. Product lead Jack Krawczyk first shared Bard with 10,000 trusted testers, pitching it as a creative helper rather than a search tool. Behind closed doors at Google, Alphabet head Sundar Pichai and Google co-founders Larry Page and Sergey Brin held crisis meetings to speed up AI work after OpenAI’s ChatGPT became a global hit [4]. Google AI chief Jeff Dean joined Pichai in weighing the risks of early rollouts, while Reuters and SemiAnalysis estimated that adding chat to Google Search could cost the firm billions.
Google announced the Gemini model family on December 6, 2023, setting a unified base for multi-task AI [4]. In February 2024, Google dropped the Bard name for good, renaming the bot Gemini and swapping out the Duet AI tag across Google Cloud and Workspace tools.
Google offers Gemini in several sizes, from light Nano builds on Android to Flash, Pro, and big Ultra setups on Vertex AI. Early public tests drew heat over flawed image outputs, leading Google to pause human image generation while staff reworked safety tools. Later updates through 2025 and 2026 cut response delay and boosted coding strength. Today, that same model family writes thousands of JavaScript tasks to beat Minecraft, showing how far Google’s autonomous tools have come since the early Bard days [4].
- PREPRINT Faldor, M., Singh, S., & Raileanu, R. (2026). A team of Gemini 4 Argon agents beats Minecraft. arXiv. [Article Link]
- ONLINE NEWS Dent, S. (2026, October 19). Google Gemini 4 agents just beat Minecraft with no human intervention. Engadget. [Article Link]
- ONLINE NEWS Bastian, M. (2026, October 18). Four Gemini 4 Argon agents beat Minecraft without any human help. The Decoder. [Article Link]
- WEBSITE Wikipedia contributors. (2026). Google Gemini. Wikipedia. [Article Link]
